The objective of this study is to address the capabilities of multi-temporal optical images to estimate the fine-scale yields variability of wheat (at a spatial scale of 30 meters). Time series of images were collected over a study site located in southwestern France throughout four successive agricultural seasons, together with intra-field yields measured by a surveying harvesting machine equipped with dGPS system on track mode. The methodology is based on the Landsat-8 and Sentinel-2 satellite images acquired after the sowing and before the harvest of the crop, the reflectance constituting the input variables of a statistical algorithm (random forest). The large dataset allows independent training and validating steps on more than one thousands of measurements, useful for testing the robustness of the proposed approaches.

The best performances are obtained when the NDVI is combined with the previous yield maps, regardless the considered agricultural season. In such case, the agricultural season 2014 shows the lower level of performances with a R² of 0.44 and a RMSE of 8.13 q.h^-1 (corresponding to a relative error of 12.9%), the three other years being associated with values of R² close or upper to 0.60 and RMSE lower than 7 q.h^-1 (corresponding to a relative error inferior to 11.3%). Such level of error on yield estimates appears acceptable, values of RMSE being lower than the observed variability, whatever the considered year (mean standard deviation of 11.8, 9.8, 10.0 and 8.9 q.h^-1 for the yields collected in 2014, 2015, 2016 and 2017, respectively).